Just over a week after introducing Granite Switch 4.1, IBM has released version 4.1.1 to demonstrate how its modular architecture allows language models to be updated similarly to traditional software. This patch release is fully backward compatible with version 4.1, requiring no migration and delivering incremental improvements without altering the base model.
The update features a refreshed Granite Speech ASR model capable of multilingual speech-to-text conversion across ten languages. This new model includes native punctuation and capitalization, ensuring transcripts are immediately ready for downstream processing without additional formatting steps. Additionally, the Decision adapter within granitelib-core has been updated to deliver improved overall performance on the Jev Decision Index test set.
IBM highlights that the speed of this release stems from its modular design; apart from training the new ASR model and Decision adapter, no other components were recomputed. Developers can simply compose the latest adapters they require and deploy them instantly. The company is currently developing tutorials and an Adapter SDK to enable developers and third-party providers to build their own adapters, iterate on versions, and ship updates with the same efficiency as software development.
In the interim, users are encouraged to test the latest models and explore the new multilingual and punctuation capabilities through the Speech-to-Decisions demo.
Sources: Hugging Face Blog.